an anisotropic a priori error analysis for a convection ...ariel/papers/bustinza_lombardi...results...
TRANSCRIPT
-
An anisotropic a priori error analysis for aconvection–dominated diffusion problem using the
HDG method
Rommel Bustinzaa,b,∗, Ariel L. Lombardic,d, Manuel Solanoa,b
aDepartamento de Ingenieŕıa Matemática, Facultad de Ciencias F́ısicas y Matemáticas,Universidad de Concepción, Casilla 160-C, Concepción, Chile
bCentro de Investigación en Ingenieŕıa Matemática (CI2MA), Universidad de Concepción,Casilla 160-C, Concepción, ChilecMember of CONICET, Argentina
dDepartamento de Matemática, Facultad de Ciencias Exactas, Ingenieŕıa y Agrimensura,Universidad Nacional de Rosario. Av. Pellegrini 250, 2000 Rosario, Argentina
Abstract
This paper deals with the a priori error analysis for a convection-dominated
diffusion 2D problem, when applying the HDG method on a family of anisotropic
triangulations. It is known that in this case, boundary or interior layers may
appear. Therefore, it is important to resolve these layers in order to recover, if
possible, the expected order of approximation. In this work, we extend the use
of HDG method on anisotropic meshes. To this end, some assumptions need to
be asked to the stabilization parameter, as well as to the family of triangulations.
In this context, when the discrete local spaces are polynomials of degree k ≥ 0,
this approach is able to recover an order of convergence k + 12 in L2 for all the
variables. Numerical examples confirm our theoretical results.
Keywords: Hybridizable Discontinuous Galerkin,
convection-dominated diffusion problem, anisotropic meshes
2010 MSC: 65N30; 65N12; 65N15
∗Corresponding authorEmail addresses: [email protected] (Rommel Bustinza),
[email protected] (Ariel L. Lombardi), [email protected] (Manuel Solano)
Preprint submitted to Elsevier November 14, 2018
-
1. Introduction
The first studies of convection diffusion problems applying discontinuous
Galerkin (DG) methods, on a shape-regular family of triangulations, are referred
to [1, 2], in the early 2000. Since then, many other DG methods have been used
for this kind of problem. For example, in [3, 4, 5, 6, 7, 8] the authors consider the
local discontinuous Galerkin (LDG) methods, while the multiscale discontinuous
Galerkin method (see [9] for an overview) is used in [10, 11]. The interior penalty
discontinuous Galerkin (IP-DG) methods are applied in [12, 13], the method
of Bauman and Oden is considered in [14], the mixed-hybrid DG method is
employed in [15] and the HDG methods are used in [16, 17, 18, 19, 20, 21].
HDG methods are a brand new class of DG schemes, that have been used
lately. We refer [22] for a description of the technique, when applies to a linear
second order elliptic equation. One of the main advantages of HDG methods,
as indicated in [22], is the fact that one just has to solve a linear system on
the skeleton of the mesh, and then recover the rest of (global) unknowns via an
element-by-element calculation.
On the other hand, the fact that the exact solution of this kind of problems
may generate layers (cf. [23, 24]), makes difficult to obtain a good approximation
of it close to the layers. This could affect the rate of convergence of the method,
and it is in general improved after the layers are resolved. This improvement
can be done by considering meshes whose elements are concentrated along the
layers. To do this, we would need to know in advance if there are layers, and
if so, where they are. Since we do not usually know the exact solution, one
alternative is the development of a suitable a posteriori error estimate which let
us to perform an adaptive procedure to the mesh in order to capture the layer.
In [25] the authors propose a reliable and efficient a posteriori error estimate for
a convection-dominated diffusion reaction problem, and include some numerical
tests that validate the good behaviour and robustness of the estimator.
Concerning anisotropic meshes, we can refer to [26], where the authors
present an LDG a priori error analysis of a 2D convection-dominated diffu-
2
-
sion problem using Shishkin quadrilateral meshes, when the exact solution
has exponential boundary layers. Our aim is to develop HDG methods for
a convection-dominated diffusion 2D problem, when considering a sequence of
simplicial meshes which may contain anisotropic elements. It is known that
anisotropic meshes should be best suited for this kind of problem.
However, in this situation, the regularity property of the meshes is no longer
valid. Instead of this, we require that the meshes satisfy the maximum angle
condition (cf. [27]). This would be the first HDG analysis in this direction, and
from certain point of view, it generalizes the 2D a priori error analysis for a larger
family of triangulations, when HDG method is applied. To this end, we follow
ideas given in [12] and [19]. We remark at least two differences of our analysis
with respect to [19]. First, the numerical vector flux we introduced is given
in the sense of LDG scheme. Secondly, despite [19], we only need to consider
the standard local L2−orthogonal projection operators and their approximation
properties on anisotropic triangles. As result, we deduce that ||u − uh||T , the
L2−norm of error u− uh in the triangulation T , satisfies
1
||ϕ||h||u − uh||L2(Ω) = O(hk+0.5) ,
once the layers have been resolved. Here, u is the exact solution, uh its HDG-
approximation, and ϕ is a suitable function that depends on T , and whose norm
can be bounded by �−1. Here, � represents the diffusion coefficient, h is intended
to be the mesh size considered to obtain uh, that is, h := max{hK : K ∈ T },
with hK being the diameter of K. Additional symbols and notations will be
properly introduced in Section 3.
The rest of the paper is organized as follows. In Section 2, we introduce
the model problem, deduce the HDG formulation and discuss on its unique
solvability. The details of the anisotropic a priori error analysis are given in
Section 3. Next, in Section 4, several numerical examples are shown, whose
results are in agreement with our convergence analysis, even in cases where the
maximum angle is close to π. Finally, we end this work, giving some conclusions
and final remarks.
3
-
2. Convection–diffusion problem
Here, we consider the model problem
�−1 q + ∇u = 0 in Ω ,
σ = q + uv in Ω ,
∇ · σ = f in Ω ,
u = g on ∂Ω ,
(1)
where Ω is a polygonal domain in R2, f ∈ L2(Ω), g ∈ H1/2(∂Ω), the convection
velocity field v ∈ [W 1,∞(Ω)]2, and has neither closed curves nor stationary
points. In addition, we require that ∇ · v ≥ 0 in Ω. This implies (see [12],
Appendix A, for a proof) that there exists a smooth function ψ so that v(x) ·
∇ψ(x) ≥ b0 ∀x ∈ Ω , for some constant b0 > 0.
Remark 2.1. When v ∈ [P0(Ω̄)]2, we can set ψ(x) := b0 v·x||v||2 .
Now, in order to define the HDG method, we let T be a triangulation of
Ω, made of triangular elements satisfying a maximum angle condition with
constant β̃ < π. This means that all the angles of the triangles in T are less or
equal than β̃ > 0 (see hypothesis M.1 in Section 3). We remind that for any
K ∈ T , hK denotes the diameter of K and h := maxK∈T hK . We also introduce
∂T := {∂K : K ∈ T }, and let E be the set of all sides F of all elements K ∈ T ,
counted once. Given K ∈ T , we denote by n the unit normal vector, exterior
to ∂K. Concerning the approximation spaces, we first introduce the space of
piecewise polynomials of degree at most k ∈ N ∪ {0}
Pk(T ) :={w ∈ L2(Ω) : w|K ∈ Pk(K) ∀K ∈ T
}.
Then, we look for the approximation of u and q in the discrete spaces Wh :=
Pk(T ) and Vh := [Pk(T )]2, respectively. We also consider the space
Mh := Pk(E) :={w ∈ L2(E) : w|F ∈ Pk(F ) ∀F ∈ E
},
4
-
for another scalar unknown that lives on the skeleton of T , well known as
numerical trace, and the afine space
Mh(g) := {µ ∈Mh : 〈µ, ζ〉F = 〈g , ζ〉F ∀ ζ ∈ Pk(F ) ∀F ∈ E ∩ ∂Ω} ,
for imposing Dirichlet boundary condition on the discrete formulation in a weak
sense. Concerning the inner products consider here, all of them are piecewise
defined. For instance,
(w, v)T :=∑K∈T
∫K
w v ∀w , v ∈ L2(Ω) ,
〈µ, ρ〉∂T :=∑K∈T
∫∂K
µρ ∀µ , ρ ∈ L2(∂T ) .
The definition of (·, ·)T for vector functions is given in analogous way. By || · ||Tand || · ||∂T we denote the norms induced by the corresponding inner products
defined above.
The HDG formulation reads as: Find (σh, qh, uh, ûh) ∈ Vh × Vh ×Wh ×Mh,
such that
(�−1 qh, r)T − (uh,∇ · r)T + 〈r · n, ûh〉∂T = 0 ∀ r ∈ Vh ,
(σh,ρ)T − (qh,ρ)T − (uh v,ρ)T = 0 ∀ρ ∈ Vh ,
−(σh,∇w)T + 〈σ̂ · n, w〉∂T = (f, w)T ∀w ∈Wh ,
〈ûh, µ〉∂Ω = 〈g, µ〉∂Ω ∀µ ∈Mh ,
〈σ̂ · n, µ〉∂T \∂Ω = 0 ∀µ ∈Mh ,
(2)
where we set σ̂ := qh + uh v + τ(uh − ûh)n on ∂T . Here we assume that
τ ∈ P0(E) is a non-negative parameter on E .
We deduce from the second equation in (2) that σh = qh +PVh(uh v) ∈ Vh,
with PVh being the L2−projection operator onto Vh.
Remark 2.2. We notice that when v ∈ [P0(Ω)]2, we have that uh v ∈ [Pk(T )]2,
provided uh ∈ Pk(T ). This allows us to set the numerical flux σ̂ := σh + τ(uh −
ûh)n on ∂T , in the same spirit of LDG method.
5
-
Then, we derive an equivalent HDG formulation, which reads: Find (qh, uh, ûh) ∈
Vh ×Wh ×Mh, such that
(�−1 qh, r)T − (uh,∇ · r)T + 〈r · n, ûh〉∂T = 0 ,
−(qh,∇w)T − (uhv,∇w)T + 〈(qh + uhv) · n+ τ(uh − ûh), w〉∂T = (f, w)T ,
〈ûh, µ〉∂Ω = 〈g, µ〉∂Ω ,
〈(qh + uhv) · n+ τ(uh − ûh), µ〉∂T \∂Ω = 0 ,(3)
for any (r, w, µ) ∈ Vh ×Wh ×Mh.
The existence and uniqueness of the solution of the HDG scheme (3) is
established next. To this end, it is important the identity
(w v,∇w)T =〈
1
2(v · n)w,w
〉∂T−(
1
2(∇ · v)w,w
)T∀w ∈ H1(T ) . (4)
Theorem 2.1. If τ + 12v · n > 0 on ∂T , then the HDG formulation (3) has
one, and only one solution.
Proof. Since the discrete scheme is linear and square, it is enough to prove that
the associated homogeneous linear system
(�−1 qh, r)T − (uh,∇ · r)T + 〈r · n, ûh〉∂T = 0 ,(5)
−(qh,∇w)T − (uhv,∇w)T + 〈(qh + uhv) · n+ τ(uh − ûh), w〉∂T = 0 ,(6)
〈ûh, µ〉∂Ω = 0 ,(7)
〈(qh + uhv) · n+ τ(uh − ûh), µ〉∂T \∂Ω = 0 ,(8)
for any (r, w, µ) ∈ Vh ×Wh ×Mh, has only the trivial solution.
First, from (7) we deduce that ûh = 0 on ∂Ω. Taking r := qh, w := uh and
µ := ûh in (5), (6) and (8), respectively, and taking into account (4), we deduce
after suitable algebraic manipulations
(�−1 qh, qh)T +
(1
2(∇ · v)uh, uh
)T
+
〈(τ +
1
2v · n
)(uh − ûh), uh − ûh
〉∂T
= 0 .
Now, since τ + 12v · n > 0 on ∂T , we deduce that uh = ûh on ∂T and qh = 0
in T . When ∇ · v > 0 a.e. Ω, we have uh = 0 in T , and then ûh = 0 on ∂T .
6
-
Otherwise, (5) implies that ∇uh = 0 in T , so uh ∈ P0(T ). Since uh = ûh on T ,
we conclude that uh ∈ P0(Ω). As uh = ûh = 0 on ∂Ω, we derive that uh = 0 in
Ω. Thus, we end the proof.
�
3. An anisotropic a priori error analysis
We adapt the technique described in [19] to our case. First, we recall and
introduce some notations and requirements on the family of triangulation.
For each K ∈ T , βK denotes the maximum interior angle of K, hK :=
maxF∈∂K
|F | and hmin,K := minF∈∂K
|F |. Given one side F of K ∈ T , F⊥ denotes
the height relative to F . We introduce the principal directions of K, denoted
by s1 and s2 (with ||s1|| = ||s2|| = 1), as the directions of the sides E1 and
E2 of K, sharing the vertex of the maximum angle of K. In addition, we
consider the standard multi-index notation α := (α1, α2) ∈ Z+0 × Z+0 , with
length |α| := α1 + α2, and define
∂α := ∂α1s1 ∂α2s2 ∀α := (α1, α2) ∈ Z
+0 × Z
+0 ,
h̃γK := hγ11,K h
γ22,K ∀ γ := (γ1, γ2) ∈ R× R .
(9)
Here, given s a unit vector, ∂s denotes the corresponding derivative operator
with respect to the direction s, while h1,K and h2,K denote the lengths of E1
and E2, respectively.
From now on, we assume that T := {Tm} is a sequence of meshes that
satisfies:
M.1 the maximum angle condition, i.e. there is 0 < β̃ < π such that βK ≤ β̃,
∀K ∈ T ,∀ T ∈ T .
We recall here that we are not assuming the shape-regularity hypothesis, so our
family of meshes may contain arbitrary anisotropic elements satisfying the max-
imum angle condition. Hereafter, we remark that C, with or without subscript
or tildes, will denote a positive constant, that is independent of �, the mesh size,
and the maximum angle of any triangle of T .
7
-
In what follows we consider that the parameter τ satisfies the following
properties, for any T ∈ T :
H.1 ∃C0 > 0 such that maxF∈∂K
τ |F =: τmaxK ≤ C0, ∀K ∈ T .
H.2 ∃C1 > 0 such that τvK := maxF∈∂K
infF
(τ +
1
2v · n
)≥ C1 min
{�
hmin,K, 1
},
∀K ∈ T .
H.3 ∃C2 > 0 such that infF
(τ +
1
2v · n
)≥ C2 max
F|v ·n|, ∀F ∈ ∂K, ∀K ∈
T .
From now on, by a . b we mean that a ≤ C b, for some positive constant
C that is independent of the mesh size and β̃.
We also need to consider the following broken Sobolev spaces
V (T ) := [H1(T )]2 , W (T ) := H1(T ) , M(E) := L2(E) .
Next, we introduce the bilinear form B : (V (T ) ×W (T ) ×M(E)) × (V (T ) ×
W (T )×M(E))→ R given by
B((q, u, λ), (r, w, µ)) := (�−1 q, r)T − (u,∇ · r)T + 〈r · n, λ〉∂T − (q + uv,∇w)T
+ 〈(q + uv) · n + τ (u− λ) , w − µ〉∂T , (10)
for any (q, u, λ) , (r, w, µ) ∈ V (T ) ×W (T ) ×M(E) . We notice that problem
(3) can be written as: Find (qh, uh, ûh) ∈ Vh ×Wh ×Mh(g) such that
B((qh, uh, ûh), (r, w, µ)) = (f, w)T + 〈g , µ〉∂Ω ∀ (r, w, µ) ∈ Vh×Wh×Mh(0) .
(11)
In what follows, we restrict ourselves, for simplicity, to the case g = 0 on ∂Ω.
Now, taking (r, w, µ) := (qh, uh, ûh) ∈ Vh ×Wh ×Mh(0), we deduce
||�−1/2qh||2T +
∥∥∥∥∥(τ +
1
2v · n
)1/2(uh − ûh)
∥∥∥∥∥2
∂T
+
∥∥∥∥∥(
1
2(∇ · v)
)1/2uh
∥∥∥∥∥2
T
= (f, uh)T .
Unfortunately, if v is divergence-free, we do not have any control of the L2−norm
of uh, by the standard energy argument. This motivates us to proceed as in [12]
8
-
(see also [19]). Then, we introduce the norm ||| · ||| : V (T )×W (T )×M(E)→
R+0 , given by
|||(r, w, µ)|||2 := ||�−1/2r||2T + ||w||2T +∑K∈T
∥∥∥∥∥(
hKhmin,K
)1/2(τ +
1
2v · n
)1/2(w − µ)
∥∥∥∥∥2
∂K
,
for any (r, w, µ) ∈ V (T )×W (T )×M(E). Next, we consider the function
ϕT := e−ψ + χ max
K∈T
(hK
hmin,K
),
where ψ is the function introduced at the beginning of Section 2, and χ is a
(suitable) positive constant at our disposal. Then, we can establish the following
result
Lemma 3.1. Let ϕ := ϕT given above, with
χ ≥ 1 + b−10 ||∇ψ||2L∞(Ω) ||e−ψ||L∞(Ω) ||v||L∞(Ω) > 0 .
Assuming that τ + 12v · n > 0 ∀F ∈ ∂K , ∀K ∈ T , there exists C > 0,
independent of � and the mesh size, but depending on b0, ψ and v, such that
B((r, w, µ), (rϕ, wϕ, µϕ)) ≥ C |||(r, w, µ)|||2 ∀ (r, w, µ) ∈ Vh ×Wh ×Mh(0) ,
where (rϕ, wϕ, µϕ) := (ϕ r, ϕw, ϕµ).
Proof. Given (r, w, µ) ∈ Vh ×Wh ×Mh(0), we have, after integrating by parts
and doing algebraic manipulations
B((r, w, µ), (rϕ, wϕ, µϕ)) = (�−1r, ϕ r)T + (w , e
−ψ∇ψ · r)T +1
2((∇ · v)w , ϕw)T
+1
2((v · ∇ψ)w , e−ψ w)T +
〈(τ +
1
2v · n
)ϕ (w − µ) , w − µ
〉∂T(12)
Since v · ∇ψ ≥ b0 > 0, ∇ · v ≥ 0, and ϕ ≥ χ maxK∈T
(hK
hmin,K
)≥ χ in Ω, we
deduce
B((r, w, µ), (rϕ, wϕ, µϕ)) ≥ χ (�−1r, r)T + (w , e−ψ∇ψ · r)T +b02
(w , e−ψ w)T
+χ
〈maxK∈T
(hK
hmin,K
)(τ +
1
2v · n
)(w − µ) , w − µ
〉∂T
(13)
9
-
Now, applying arithmetic-geometric Cauchy-Schwarz inequality, we have, for
any δ > 0,
2∣∣(w, e−ψ∇ψ · r)T ∣∣ = 2 ∣∣∣(e−ψ/2 w, e−ψ/2∇ψ · r)T ∣∣∣
≤ δ−1(e−ψ/2∇ψ · r, e−ψ/2∇ψ · r)T ,+ δ (e−ψ/2 w, e−ψ/2 w)T
≤ δ−1||∇ψ||2L∞(Ω) (e−ψr, r)T + δ (e
−ψ w,w)T
≤ δ−1||∇ψ||2L∞(Ω) ||e−ψ||L∞(Ω) ||v||L∞(Ω) (�−1r, r)T + δ (e−ψ w,w)T ,
since 1 � �−1 ||v||L∞(Ω). This lets us to derive, from (13), that
B((r, w, µ), (rϕ, wϕ, µϕ)) ≥(χ − δ
−1
2||∇ψ||2L∞(Ω)||e
−ψ||L∞(Ω) ||v||L∞(Ω))
(�−1r, r)T
+
(b02− δ
2
)(w , e−ψ w)T + χ
〈maxK∈T
(hK
hmin,K
)(τ +
1
2v · n
)(w − µ) , w − µ
〉∂T
.
(14)
Choosing δ := b0/2 > 0 in (14), and taking into account the hypothesis on χ,
we obtain
B((r, w, µ), (rϕ, wϕ, µϕ)) ≥ (�−1r, r)T +b04
(e−ψ w , w)T
+χ
〈maxK∈T
(hK
hmin,K
)(τ +
1
2v · n
)(w − µ) , w − µ
〉∂T
≥ C |||(r, w, µ)|||2 , (15)
with C > 0 depending on b0, e−ψ, ∇ψ and v . �
We notice that the test function (r, w, µ) := (rϕ, wϕ, µϕ) in Lemma 16 does
not belong to the discrete space Vh ×Wh ×Mh(0). In order to derive our a
priori error estimate, we introduce the standard L2−projection operators ΠV ,
ΠW and PM onto Vh, Wh and Mh, respectively.
Another tool we need for the a priori error analysis, is the averaged Taylor
operator Qk of degree k ≥ 0, introduced and analyzed in [28]. Indeed, given
u ∈ Hk+1(K), we define Qku ∈ Pk(K) as
Qku(x) :=1
|K|
∫K
Tku(y, x) dy ,
10
-
with
Tku(y, x) :=∑|α|≤k
∂αu(y)(y − x)α
α!.
The approximation properties of Qk are described next.
Lemma 3.2. For any K ∈ T and any w ∈ Hk+1(K), there exists C > 0,
independent of the maximum angle βK , such that for any side F of K with
corresponding direction vector s
||w −Qkw||L2(K) ≤ C∑
|α|=k+1
h̃αK ||∂αw||L2(K) , (16)
|F |‖∂sQkw‖L2(K) ≤ C ‖w‖L2(K) , (17)
|F |‖∂s(w −Qkw)‖L2(K) ≤ C∑
|α|=k+1
h̃αK‖∂αw‖L2(K) , (18)
with h̃αK being defined as in (9).
Proof. These inequalities are obtained from [28], by rescaling arguments to a
reference element. We omit further details. �
Next, we establish a geometric relation valid on any triangle.
Lemma 3.3. For any triangle K there holds
|F⊥| ≥ 12
sin(βK)hmin,K ∀F ∈ ∂K .
Proof. Let K be a triangle, and Fa, Fb and Fc its sides such that |Fa| ≤ |Fb| ≤
|Fc|. It is enough to prove the property for the height of K relative to its largest
side. Then, we have
|F⊥c | |Fc| = 2 |K| = |Fa| |Fb| sin(βK) ,
with βK denoting the maximum angle of K. The proof follows using the fact
that
2 |Fb| ≥ |Fb| + |Fa| > |Fc| .
We omit further details. �
In addition, we also need an anisotropic version of the trace inequality, which
can be proven by standard rescaling arguments. We remark that in Lemma 2.3
11
-
in [29], it has been established an anisotropic trace inequality on tetrahedra,
applying this kind of argument.
Lemma 3.4. For any triangle K ∈ T , there exists C > 0, independent of the
mesh size and the maximum angle of K, such that for any side F of K, there
holds
||w||L2(F ) ≤ C |F⊥|−1/2(||w||L2(K) +
∑E∈∂K
|E| ||∂sEw||L2(K)
), ∀w ∈ H1(K) ,
where for any side E of K, sE represents its unit direction vector.
The next result is a consequence of estimates for averaged Taylor operator.
Lemma 3.5. Assume that w ∈ H lw+1(K), for lw ∈ [0, k] on an element K ∈ T .
Then there exists C > 0, independent of the mesh size and the maximum angle
βK , such that
||ΠWw − w||L2(K) ≤ C∑
|α|=lw+1
h̃αK ||∂αw||L2(K) .
Also, if r ∈ [H lr+1(K)]2, for lr ∈ [0, k] on an element K ∈ T . Then there exists
C > 0, independent of the mesh size and the maximum angle βK , such that
||ΠV r − r||[L2(K)]2 ≤ C∑
|α|=lr+1
h̃αK ||∂αr||L2(K)]2 ,
with ∂α and h̃αK being introduced in (9).
Proof. First, we consider the averaged Taylor approximation of w, Qkw ∈
Pk(K). Then, after noting that ΠW (Qkw) = Qkw in K, we have
||ΠWw−w||L2(K) ≤ ||ΠW (w−Qkw)||L2(K) + ||Qkw−w||L2(K) ≤ 2 ||Qkw−w||L2(K) .
The conclusion follows after applying (16). The second approximation property
is proved in analogous way. �
Lemma 3.6. Let K ∈ T and φ ∈ C1(K̄)∩W k+1,∞(K). Then, for any (r, w) ∈
[Pk(K)]2 × Pk(K) and ζ ∈ R, there exists C > 0, such that
‖(ΠV − I) (φ + ζ) r‖[L2(K)]2 ≤ ChK
sin(βK)||φ||Wk+1,∞(K) ||r||[L2(K)]2 ,
12
-
‖(ΠV − I) (φ + ζ) r‖[L2(F )]2 ≤ Ch
1/2K
sin(βK)
(hK
hmin,K
)1/2||φ||Wk+1,∞(K) ||r||[L2(K)]2 ∀F ∈ ∂K ,
‖(ΠW − I) (φ + ζ)w‖L2(K) ≤ ChK
sin(βK)||φ||Wk+1,∞(K) ||w||L2(K) ,
‖(ΠW − I) (φ + ζ)w‖L2(F ) ≤ Ch
1/2K
sin(βK)
(hK
hmin,K
)1/2||φ||Wk+1,∞(K) ||w||L2(K) ∀F ∈ ∂K .
Proof. Since (ΠV − I)((φ + ζ)r)|K = (ΠV − I)(φr)|K , and applying Lemma
3.5, we have
‖(ΠV − I)(φ+ ζ)r‖[L2(K)]2 ≤ C∑
|α|=k+1
h̃αK‖∂α(φr)‖[L2(K)]2
≤ C∑
|α|=k+1
h̃αK∑β≤α
‖∂α−βφ‖L∞(K)‖∂βr‖[L2(K)]2 ,(19)
with the constant C from Lemma 3.5. Applying the inverse inequality
‖∂βr‖[L2(K)]2 ≤ h̃−βK ‖r‖[L2(K)]2 ,
and taking into account that r is of degree less or equal than k, we obtain
‖(ΠV − I)(φ+ ζ)r‖[L2(K)]2 ≤ C̃ csc(βK) ‖φ‖Wk+1,∞(K)∑
|α|=k+1
∑β≤α,|β|≤k
h̃α−βK ‖r‖[L2(K)]2
≤ C̃ csc(βK)hK‖φ‖Wk+1,∞(K)‖r‖[L2(K)]2 ,
with C̃ > 0 independent of the mesh size and the maximum angle of K. This
concludes the proof of the first inequality.
To establish the second inequality, we take into account Lemma 3.4. Then,
we obtain
‖(ΠV − I)(φ+ ζ)r‖[L2(F )]2 = ‖(ΠV − I)(φr)‖[L2(F )]2
≤ C |F⊥|−1/2(‖(ΠV − I)(φr)‖[L2(K)]2 +
∑E∈∂K
|E|‖∂sE (ΠV − I)(φr)‖[L2(K)]2).
(20)
Introducing now the averaged Taylor k-degree polynomial Qk(φr) of φr on K
13
-
and applying Lemma 3.2, we have
|E|‖∂s(ΠV − I)(φr)‖[L2(K)]2 = |E|‖∂s(ΠV −Qk +Qk − I)(φr)‖[L2(K)]2
≤ |E|‖∂s [Qk(ΠV − I)(φr)] ‖[L2(K)]2 + |E|‖∂sE (Qk − I)(φr)‖[L2(K)]2
≤ C
‖(ΠV − I)(φr)‖[L2(K)]2 + ∑|α|=k+1
h̃αK‖∂α(φr)‖[L2(K)]2
.(21)
Now, by the same argument used in equation (19) and from the first inequality
of this Lemma, we deduce from (21)
|E|‖∂sE (ΠV − I)(φr)‖[L2(K)]2 ≤ C csc(βK)hK‖φ‖Wk+1,∞(K)‖r‖[L2(K)]2 ,
and then, after replacing back in (20) and taking into account Lemma 3.3, we
conclude that
‖(ΠV − I)(φ+ ζ)r‖[L2(F )]2 ≤ C csc(βK)h−1/2min,K hK‖φ‖Wk+1,∞(K)‖r‖[L2(K)]2 ,
with C > 0 independent of the mesh size and the maximum angle βK .
Third and fourth inequalities are proved analogously. �
In what follows, we set ||ϕ||h := maxK∈T
||ϕ||Wk+1,∞(K), and Dβ̃ := maxK∈T csc(βK).
Lemma 3.7. There exists h0 > 0, independent of �, but dependent of β̃, so that
for any h < h0, there holds the following inf-sup condition: There exists C > 0,
independent of �, the maximum angle of all K ∈ T and the mesh size, such that
for any (q, u, λ) ∈ Vh ×Wh ×Mh(0)
sup(r,w,µ)∈Vh×Wh×Mh(0)
(r,w,µ) 6=(0,0,0)
B((q, u, λ), (r, w, µ))
|||(r, w, µ)|||≥ C
Dβ̃ ||ϕ||h|||(q,u, λ)||| (22)
Proof. First, we let (q, u, λ) ∈ Vh ×Wh ×Mh(0), and introduce δqϕ := (I −
ΠV )qϕ ∈ V (T ), δuϕ := (I−ΠW )uϕ ∈W (T ) and δλϕ := (I−PM )λϕ ∈M(E).
Then, we have
B((q, u, λ), (δqϕ, δuϕ, δλϕ)) = (�−1 q, δqϕ)T − (u,∇ · δqϕ)T + 〈δqϕ · n , λ〉∂T
14
-
−(q + uv,∇ δuϕ)T + 〈(q + uv) · n + τ(u− λ) , δuϕ〉∂T
−〈(q + uv) · n + τ(u− λ) , δλϕ〉∂T
= (�−1 q , δqϕ)T + 〈δqϕ · n , λ − u〉∂T + 〈τ(u− λ) , δuϕ〉∂T .
Now, our aim is to bound each one of the three terms above. Applying Cauchy-
Schwarz inequality and first approximation property in Lemma 3.6, we have
(�−1q, δqϕ)K ≤ ||�−1/2q||[L2(K)]2 ||�−1/2δqϕ||[L2(K)]2 . csc(βK)hK ||�−1/2q||2[L2(K)]2 .
On the other hand, since τvK ≤ τ + 12v · n on ∂K, and taking into account
second approximation property in Lemma 17, we derive
〈δqϕ · n , λ − u〉∂K ≤ ||(τvK)−1/2δqϕ · n||L2(∂K) ||(τvK)1/2(λ − u)||L2(∂K)
≤(�
τvK
)1/2||�−1/2δqϕ||L2(∂K)
∥∥∥∥∥(τ +
1
2v · n
)1/2(λ − u)
∥∥∥∥∥L2(∂K)
. csc(βK)h1/2K
(�
τvK
)1/2||�−1/2q||[L2(K)]2
∥∥∥∥∥(
hKhmin,K
)1/2(τ +
1
2v · n
)1/2(λ − u)
∥∥∥∥∥L2(∂K)
. csc(βK) (h2K + � hK)
1/2
∥∥∥∥∥(
hKhmin,K
)1/2(τ +
1
2v · n
)1/2(λ − u)
∥∥∥∥∥L2(∂K)
||�−1/2q||[L2(K)]2 .
In addition, considering H.3, it is not difficult to check
〈τ(u− λ) , δuϕ〉∂K =〈(
τ +1
2v · n
)(u− λ) , δuϕ
〉∂K
−〈
1
2(v · n)(u− λ) , δuϕ
〉∂K
≤
∥∥∥∥∥(τ +
1
2v · n
)1/2(u− λ)
∥∥∥∥∥L2(∂K)
+
∥∥∥∥∥∣∣∣∣12v · n
∣∣∣∣1/2 (u− λ)∥∥∥∥∥L2(∂K)
||δuϕ||L2(∂K). csc(βK)h
1/2K
∥∥∥∥∥(
hKhmin,K
)1/2(τ +
1
2v · n
)1/2(u− λ)
∥∥∥∥∥L2(∂K)
||u||L2(K) .
15
-
Then, we have
B((q, u, λ), (δqϕ, δuϕ, δλϕ)) .∑K∈T
csc(βK)hK ||�−1/2q||2[L2(K)]2
+∑K∈T
csc(βK) (h2K + � hK)
1/2
∥∥∥∥∥(
hKhmin,K
)1/2(τ +
1
2v · n
)1/2(λ − u)
∥∥∥∥∥L2(∂K)
||�−1/2q||[L2(K)]2
+∑K∈T
csc(βK)h1/2K
∥∥∥∥∥(
hKhmin,K
)1/2(τ +
1
2v · n
)1/2(u− λ)
∥∥∥∥∥L2(∂K)
||u||L2(K)
. Dβ̃ h1/2 |||(q, u, λ)|||2 .
Thanks to Lemma 3.1, we deduce there exists Ĉ > 0 independent of the mesh
size and � such that
B((q, u, λ), (δqϕ, δuϕ, δλϕ)) ≤ Ĉ Dβ̃ h1/2B((q, u, λ), (qϕ, uϕ, λϕ)) .
Then, we conclude that there exists h0 > 0 such that for any h < h0 there holds
B((q, u, λ), (δqϕ, δuϕ, δλϕ)) ≤1
2B((q, u, λ), (qϕ, uϕ, λϕ)) ,
from which is inferred that (applying again Lemma 3.1)
B((q, u, λ), (ΠV qϕ,ΠWuϕ, PMλϕ)) ≥1
2B((q, u, λ), (qϕ, uϕ, λϕ)) ≥
C
2|||(q, u, λ)|||2 .
Now, applying triangle inequality, Lemma 3.6, we also show that for any h < h0,
there holds
|||(ΠV qϕ,ΠWuϕ, PMλϕ)||| . Dβ̃ ||ϕ||h |||(q, u, λ)||| ,
which let us to conclude the desired result. �
Now, we let (q, u) be the exact solution, and (qh, uh, ûh) ∈ Vh×Wh×Mh(0)
the solution of (11). It is not difficult to check that (11) is consistent with the
exact solution, which means
B((q, u, u), (r, w, λ)) = (f, w)T ∀ (r, w, λ) ∈ Vh ×Wh ×Mh(0) .
This yields to the orthogonality relation
B((q−qh, u−uh, u− ûh), (r, w, λ)) = 0 ∀ (r, w, λ) ∈ Vh×Wh×Mh(0) . (23)
16
-
We introduce now
eqh := qh − ΠV q , δq := q − ΠV q ,
euh := uh − ΠWu , δu := u − ΠWu ,
eûh := ûh − PMu , δû := u − PMu .
Thanks to (23) and definition of projections, we deduce the following identity
Lemma 3.8. For any (r, w, µ) ∈ Vh ×Wh ×Mh(0), there holds
B((eqh, euh, e
ûh), (r, w, µ)) = (�
−1 δq, r)T + 〈δq · n , w − µ〉∂T
+ 〈(τ + v · n)δu , w − µ〉∂T − 〈τ δû , w − µ〉∂T .(24)
Finally, we can prove the main result of this paper. We recall that ∂α and
h̃αK have been introduced in (9).
Theorem 3.1. For h < h0 (introduced in Lemma 3.7), there exists C > 0,
independent of mesh size and parameter �, such that
1
||ϕ||h|||(q − qh, u − uh, u − ûh)||| ≤ C Dβ̃
∑K∈T
�1/2 ∑|α|=lq+1
h̃αK ||∂α∇u||[L2(K)]2
+ �∑|α|=lq|β|=1
h̃α+0.5βK ||∂α+β∇u||[L2(K)]2 +
∑|α|=lu+1
h̃αK ||∂αu||L2(K) +∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K)
.Proof. First, we bound each term on the right hand side in (24). We have, for
each K ∈ T
(�−1δq, r)K ≤ ||�−1/2 δq||[L2(K)]2 ||�−1/2 r||[L2(K)]2 .
17
-
〈δq · n , w − µ〉∂K = 〈(τvK)−1/2δq · n , (τvK)1/2(w − µ)〉∂K
≤(�
τvK
)1/2||�−1/2 δq||[L2(∂K)]2
∥∥∥∥∥(τ +
1
2v · n
)1/2(w − µ)
∥∥∥∥∥L2(∂K)
〈(τ + v · n) δu , w − µ〉∂K =〈(
τ +1
2v · n
)δu , w − µ
〉∂K
+
〈1
2(v · n) δu , w − µ
〉∂K
.
∥∥∥∥∥(τ +
1
2v · n
)1/2(w − µ)
∥∥∥∥∥L2(∂K)
||δu||L2(∂K) ,
〈τ δû , w − µ〉∂K =〈(
τ +1
2v · n
)δû , w − µ
〉∂K
− 12〈(v · n) δû , w − µ〉∂K
.
∥∥∥∥∥(τ +
1
2v · n
)1/2(w − µ)
∥∥∥∥∥L2(∂K)
||δû||L2(∂K) .
Next, we take into account the approximation results (Lemmas 3.5 and 3.4)
||�−1/2δq||[L2(K)]2 . �−1/2 csc(βK)∑
|α|=lq+1
h̃αK ||∂αq||[L2(K)]2 ,
||�−1/2δq||[L2(∂K)]2 . �−1/2 csc(βK)(
hKhmin,K
)1/2 ∑|α|=lq|β|=1
h̃α+0.5βK ||∂α+βq||[L2(K)]2 ,
||δu||L2(∂K) . csc(βK)(
hKhmin,K
)1/2 ∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K) ,
||δû||L2(∂K) . csc(βK)(
hKhmin,K
)1/2 ∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K) ,
where lq , lu ∈ [0, k], and C a positive constant that does not depend on the
maximum angle βK . Then, for any (r, w, µ) ∈ Vh ×Wh ×Mh(0), we deduce
18
-
B((eqh, euh, e
ûh), (r, w, µ)) .
∑K∈T
csc(βK)
∑|α|=lq|β|=1
h̃α+0.5βK ||∂α+βq||[L2(K)]2
+ �−1/2∑
|α|=lq+1
h̃αK ||∂αq||[L2(K)]2 +∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K)
|||(r, w, µ)||| .Since (eqh, e
uh, e
ûh) ∈ Vh×Wh×Mh(0), we apply Lemma 3.7, and then for h small
enough we have
1
||ϕ||h|||(eqh, e
uh, e
ûh)||| . Dβ̃
∑K∈T
∑|α|=lq|β|=1
h̃α+0.5βK ||∂α+βq||[L2(K)]2
+ �−1/2∑
|α|=lq+1
h̃αK ||∂αq||[L2(K)]2 +∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K)
.(25)
As q = �∇u, we derive
1
||ϕ||h|||(eqh, e
uh, e
ûh)||| . Dβ̃
∑K∈T
� ∑|α|=lq|β|=1
h̃α+0.5βK ||∂α+β∇u||[L2(K)]2
+ �1/2∑
|α|=lq+1
h̃αK ||∂α∇u||[L2(K)]2 +∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K)
.By approximation properties of the projection, we also deduce
1
||ϕ||h|||(δq, δu, δû)||| . Dβ̃
∑K∈T
�1/2 ∑|α|=lq+1
h̃αK ||∂α∇u||[L2(K)]2
+∑
|α|=lu+1
h̃αK ||∂αu||L2(K) +∑|α|=lu|β|=1
h̃α+0.5βK ||∂α+βu||L2(K)
.Finally, applying triangle inequality, we derive the result and conclude the proof.
�
19
-
Remark 3.1. When, in addition, � . hmin,K ∀K ∈ T , and the regularity of
u and q are such that lu = k and lq = max{0, k − 1} respectively, we obtain
1
||ϕ||h|||(q − qh, u − uh, u − ûh)||| = O(hk+0.5) .
Otherwise, the above expression would behave as O(hr), with r ∈ [k − 1/2, k +
1/2], and makes sense for k > 0.
Remark 3.2. The current analysis requires the maximum angle condition, which
allows us to bound Dβ̃ uniformly in our main a priori result (cf Theorem 3.1).
Otherwise, this constant Dβ̃ could blow up as the maximum angle is closer to π.
Remark 3.3. In order to obtain an error estimate of σ − σh, we take into
account the local inequality
||σ − σh||[L2(K)]2 ≤ �1/2 ||�−1/2(q − qh)||[L2(K)]2 + ||v||[L∞(K)]2 ||u− uh||L2(K)
+∑
|α|=lr+1
h̃αK ||∂α(uv)||[L2(K)]2 ∀K ∈ T .
Remark 3.4. Our main result, given in Theorem 3.1, is valid also when we
consider the numerical flux introduced in [19]
σ̂h := qh + ûh v + τ(uh − ûh) on ∂T .
In this case, we need to assume that the parameter τ is defined such that τ −12v ·n > 0 on ∂T . This should be taken into account to define the corresponding
norms, the analogous properties H.1-H.3 for τ , etc.
4. Numerical results
In the following examples the stabilization parameter τ in each edge e is
taken as τe = τed + τec , where τ
ec := supx∈e |v(x) · n| and τed := min(�/he, 1).
We compute the errors eq := ‖�−1/2(q − qh)‖[L2(Ω)]2 , eu := ‖u − uh‖L2(Ω),
eσ := ‖σ − σh‖[L2(Ω)]2 ,
eû : =
(∑K∈T
∥∥∥∥( hKhmin,K)1/2(
τ +1
2v · n
)1/2(uh − ûh)
∥∥∥∥2L2(∂K)
)1/2.
20
-
On the other hand, the estimate provided in Theorem 3.1 depends on ‖ϕ‖h,
which depends on the triangulation T and verifies
χ maxK∈T
(hK
hmin,K
)≤ ‖ϕ‖h ≤ ‖e−ψ‖Wk+1,∞(Ω) + χ max
K∈T
(hK
hmin,K
).
Since χ and ‖e−ψ‖Wk+1,∞(Ω) are independent of the mesh, the quantity MT :=
maxK∈T
(hK
hmin,K
)is a suitable indicator of the behavior of ‖ϕ‖h. Based on this ob-
servation, for each variable, we compute the experimental order of convergence
(e.o.c.) as
e.o.c. = log
(eT1/MT1eT2/MT2
)/log(hT1/hT2),
where eT1 and eT2 are the errors associated to the corresponding variable con-
sidering two consecutive meshsizes hT1 and hT2 , respectively.
4.1. Unstructured meshes
In this section we show the results obtained using anisotropic unstructured
meshes. We use BAMG ([30]) to generate an initial anisotropic mesh. Since
the goal of this work is to show the performance of the HDG method, in all the
example the exact solution is known. Hence, BAMG creates the mesh based on
a metric tensor that involves the Hessian of the solution. Then, we uniformly
refine this initial mesh by dividing the triangles by the midpoints of the edges.
This procedure keeps the anisotropy of the mesh, preserves angles and MT is
the same for every mesh.
4.1.1. Boundary layers
Example 1. We consider the domain Ω =]0, 1[2 and velocity v = (1, 1)t.
The exact solution is taken to be u(x, y) = xy(1− e�−1(x−1))(1− e�−1(y−1))
(1− e−�−1)(1− e−�−1)−
sin(3xπ/2)− sin(3yπ/2) + 2. It has boundary layers at {x = 1} and {y = 1} for
small values of �. Here, we have added sinusoidal terms so that, away from the
boundary layers, the solution does not behave as a quadratic function when �
is small. This will allows us to study the convergence rates for k > 1. In this
21
-
first example we set � = 10−3. In Figure 1 we show the initial mesh and a zoom
of it at the top-right corner. Figure 2 displays the approximate solution uh for
k = 1 and k = 2 considering a uniform mesh (left) and the anisotropic mesh
(right) showed in Figure 1. We clearly observe that the uniform mesh does not
resolve the boundary layer, but if that suitable anisotropic mesh is considered,
the approximation does not exhibit oscillations near the layers. In this case, all
the meshes satisfy maxK∈T
(hK
hmin,K
)1/2= 7.14 and max
K∈TβK = 179.5394
◦. Table 1
shows the history of convergence of the method which agrees with Remark 3.1
since the solution is smooth and maxK∈T
(hK
hmin,K
)1/2is bounded. In some cases
(k = 0, 1 and 2) the order of convergence of u is higher than expected. We point
out that since the maximum angle of these meshes is close to π, the constant
Dβ̃ in Theorem 3.1 is big. Even though, errors in Table 1 are small.
On the other hand, we numerically study the spectral condition number
κ of the global matrix associated to ûh. We recall that, for pure diffusion
problems, [31] reported numerical experiments indicating that κ is proportional
to (k + 1)h−2. For the case of convection-dominated diffusion problems, [15]
proved that κ behaves as h−2 on isotropic meshes. In Table 2 we display the
experimental order (e.o.) such that κ is proportional to h(e.o.). We observe that
it is close to -2, which agrees with the results presented in [15, 31].
Example 2. We consider the same squared domain as previous example and
the exact solution
u(x, y) = x2(y(1− y) + e−
y√� + e
− (1−y)√�),
with � = 10−3. We take v = (1, 0)t. This solution has two boundary layer on the
horizontal axis. The initial mesh is displayed in Figure 3 (left) and uh consid-
ering k = 1 and N = 7824 is shown on the right. The history of convergence of
the method provides similar conclusions as in previous example, hence we omit
the corresponding table. Here, all the meshes satisfy maxK∈T
(hK
hmin,K
)1/2= 17.87
and maxK∈T
βK = 179.6203◦.
22
-
Figure 1: Example 1: Initial mesh with N = 823 (left) and a zoom-in on the upper-right
corner (right).
4.1.2. Interior layer
Example 3. Let us consider the same domain as before and the exact solution
u(x, y) =1
1 + e−(x+y−0.8)
5�
with � = 10−3. It has an interior layer along the segment described by x +
y = 0.8. We take v = (2, 3)t. In Figure 4 (left) we display the initial mesh
generated with BAMG. Its corresponding approximated solution uh (k = 1) is
depicted on the right. No oscillation are observed near the interior layer since
the initial mesh is fine enough in that region. In this case, all the meshes satisfy
maxK∈T
(hK
hmin,K
)1/2= 18.27 and max
K∈TβK = 179.7651
◦. Once again, in accordance
with Remark 3.1, the order of convergence for eq and eû seems to be at least
hk+0.5. Moreover, for k = 0, 1 and 2 the order of convergence for eu is higher
than expected.
4.1.3. Non constant convection
Example 4. We consider the same exact solution as in Example 3 but consid-
ering a non-constant convective field v = (u, u)t. We do not displays the results
since they provide similar conclusion to the ones obtained in Example 3.
23
-
Figure 2: Approximate solution uh of Example 1 considering k = 1 (top row) and k = 2
(bottom row) when � = 10−3. Left column: uniform mesh with N = 882 elements. Top-left:
anisotropic mesh N = 823 elements.
4.2. Shishkin meshes
Example 5. We choose the same problem as in Example 1, but considering
a Shishkin mesh (Figure 5) constructed as follows (we refer to Section 2.4.2 in
[32]). We set a = min {0.5, (k + 1)� log(M)}, for a given an integer M . Then,
the intervals [0, 1 − a] and [1 − a, 1] (on both axes) are uniformly divided in
M subintervals. Finally, each rectangle on this grid is divided in two triangles
with hypotenuse parallel to the vector v = (1, 1)t. Because of this construction,
maxK∈T
βK = 90◦ and then Dβ̃ = 1. Moreover, MT decreases with k and h as we
can deduce from Table 4. Moreover, � is always greater than hmin,K for all K
in the triangulation T . Hence, Remark 3.1 predicts an order of convergence of
hr with r ∈ [k − 1/2, k + 1/2].
24
-
k N eq e.o.c eu e.o.c eσ e.o.c eû e.o.c
0 823 2.81e-01 −− 2.72e-01 −− 3.85e-01 −− 1.92e+00 −−
3292 1.87e-01 0.59 1.38e-01 0.98 1.95e-01 0.98 1.39e+00 0.46
13168 1.25e-01 0.58 7.07e-02 0.96 1.00e-01 0.96 1.02e+00 0.45
52672 8.62e-02 0.54 3.71e-02 0.93 5.24e-02 0.93 7.50e-01 0.44
210688 6.06e-02 0.51 2.02e-02 0.88 2.85e-02 0.88 5.59e-01 0.42
1 823 3.30e-02 − 2.44e-02 −− 3.45e-02 −− 2.26e-01 −−
3292 1.05e-02 1.66 5.86e-03 2.06 8.28e-03 2.06 8.45e-02 1.42
13168 4.03e-03 1.38 1.39e-03 2.08 1.95e-03 2.09 3.10e-02 1.45
52672 1.61e-03 1.32 3.22e-04 2.10 4.48e-04 2.12 1.12e-02 1.47
210688 6.61e-04 1.29 7.25e-05 2.15 9.68e-05 2.21 4.08e-03 1.46
2 823 4.85e-03 −− 1.90e-03 −− 2.68e-03 −− 2.66e-02 −−
3292 9.84e-04 2.30 2.21e-04 3.10 3.16e-04 3.09 5.82e-03 2.19
13168 1.61e-04 2.61 2.60e-05 3.09 3.74e-05 3.08 1.28e-03 2.19
52672 3.05e-05 2.40 3.03e-06 3.10 4.37e-06 3.10 2.63e-04 2.28
210688 6.20e-06 2.30 3.55e-07 3.10 4.98e-07 3.13 5.21e-05 2.34
3 823 1.37e-03 −− 1.57e-04 −− 2.28e-04 −− 4.42e-03 −−
3292 2.18e-04 2.66 1.13e-05 3.79 2.05e-05 3.48 8.59e-04 2.36
13168 2.18e-05 3.32 9.29e-07 3.61 1.91e-06 3.43 1.26e-04 2.77
52672 2.03e-06 3.42 8.39e-08 3.47 1.87e-07 3.35 1.51e-05 3.06
210688 1.30e-07 3.97 7.09e-09 3.56 1.51e-08 3.63 1.64e-06 3.20
Table 1: History of convergence of Example 1.
In Table 5 we display the history of convergence of the method considering
� = 10−3. First of all, when k = 0, we observe no convergence for qh which
agrees with Remark 3.1 because it does not guarantee convergence for this case.
Even though the error eq decreases when N increases, MT decreases faster for
N ≥ 2048 which explains the negative value for the e.o.c. The experimental
convergence rate for eu and eû seems to behave as expected, i.e., O(hr) with
25
-
k = 1 k = 2 k = 3
N κ e.o. κ e.o. κ e.o.
823 7.18e+04 − 1.17e+05 − 2.00e+05 −
3292 4.45e+05 −2.63 7.21e+05 −2.62 1.16e+06 −2.53
13168 1.70e+06 −1.94 2.74e+06 −1.93 4.42e+06 −1.93
52672 6.64e+06 −1.97 1.07e+07 −1.96 1.72e+07 −1.96
210688 2.62e+07 −1.98 4.23e+07 −1.98 6.81e+07 −1.98
Table 2: Condition number (κ) of the global matrix of Example 1.
Figure 3: Example 2: Initial mesh with N = 489 (left). Approximate solution uh with
N = 7824, k = 1 and � = 10−3 (right).
r ≥ 0.5. On the other hand, when k > 0, all the variables converge with order
in the range predicted by Remark 3.1. Except that, the convergence rate for eu
26
-
Figure 4: Example 3: Initial mesh with N = 1020 (left) and corresponding approximated
solution uh (k = 1).
is a bit higher than expected when k = 1.
We consider now � = 10−9. As we see in Table 6, MT is higher than the one
corresponding to previous case but also decreases with N and k. According to
Table 7, the experimental rate of convergence of all the variables agrees with the
range predicted by Remark 3.1. Once again, uh converges to u with an order
a bit higher than hk+0.5. The error eq in the last mesh and k = 3 is probably
being affected by round-off errors. In fact, eq has been weighted by �−1/2 which
in this case is 104.5, i.e., without this weight the error would be of order 10−12.
27
-
k N eq e.o.c eu e.o.c eσ e.o.c eû e.o.c
0 1020 7.02e-04 −− 9.46e-04 −− 3.41e-03 −− 7.69e-01 −−
4080 4.72e-04 0.57 4.87e-04 0.96 1.75e-03 0.96 5.45e-01 0.50
16320 3.75e-04 0.33 2.60e-04 0.91 9.35e-04 0.91 3.86e-01 0.50
65280 3.42e-04 0.13 1.41e-04 0.88 5.08e-04 0.88 2.74e-01 0.49
261120 3.28e-04 0.06 8.14e-05 0.80 2.93e-04 0.80 1.96e-01 0.49
1 1020 4.27e-05 −− 2.05e-05 −− 7.40e-05 −− 8.11e-03 −−
4080 1.37e-05 1.64 3.91e-06 2.39 1.41e-05 2.39 2.89e-03 1.49
16320 4.05e-06 1.76 5.95e-07 2.72 2.16e-06 2.71 1.03e-03 1.49
65280 1.19e-06 1.77 1.20e-07 2.31 4.35e-07 2.31 3.65e-04 1.49
261120 3.92e-07 1.60 3.08e-08 1.96 1.11e-07 1.97 1.30e-04 1.49
2 1020 5.61e-06 −− 4.45e-06 −− 1.60e-05 −− 3.29e-04 −−
4080 1.71e-06 1.72 3.76e-07 3.57 1.34e-06 3.58 7.11e-05 2.21
16320 4.32e-07 1.98 4.68e-08 3.01 1.55e-07 3.11 1.57e-05 2.18
65280 6.75e-08 2.68 5.40e-09 3.12 1.58e-08 3.30 3.28e-06 2.26
261120 9.51e-09 2.83 5.32e-10 3.34 2.54e-09 2.64 6.23e-07 2.40
3 1020 2.65e-06 −− 1.09e-06 −− 3.92e-06 −− 4.35e-05 −−
4080 7.85e-07 1.76 8.76e-08 3.63 2.97e-07 3.72 1.01e-05 2.11
16320 1.19e-07 2.72 1.26e-08 2.80 4.12e-08 2.85 2.07e-06 2.28
65280 1.10e-08 3.44 1.12e-09 3.49 3.52e-09 3.55 2.76e-07 2.90
261120 7.13e-10 3.94 7.13e-11 3.97 2.10e-10 4.07 2.82e-08 3.29
Table 3: History of convergence of Example 3.
5. Conclusions and final comments
In this work we have developed an a priori error analysis for the convection
dominated diffusion problem in 2D, when using the HDG method on a family
of anisotropic triangulations. We adapt ideas given in [12] and [19], in order
to follow the dependence of the constants on the diffusion coefficient �, and in
this case also on the uniform bound of the maximum angle of the triangulation.
28
-
Figure 5: Shishkin mesh for example 1 with M = 3, k = 1 and � = 10−3 (left) and a zoom on
the upper right corner (right).
As result, we deduce that when � is small enough, the corresponding rate of
convergence is O(hk+0.5). Otherwise, the rate of convergence would behave as
O(hr), for some r ∈ [k − 0.5, k + 0.5]. Numerical examples are in agreement
with our theoretical results, even when the maximum angle on the family of
triangulations is very close to π. Since the a priori error analysis require usual
L2−projectors and does not need the use of Lagrange interpolation operator,
we believe that the analysis can be adapted for 3D case (anisotropic tetrahedra)
in a natural way. On the other hand, as the presence of boundary or inner
layers is very natural in this kind of problems, it would be better to count with
29
-
N k = 0 k = 1 k = 2 k = 3
32 3.78e+01 2.68e+01 2.19e+01 1.90e+01
128 2.68e+01 1.90e+01 1.55e+01 1.34e+01
512 2.19e+01 1.55e+01 1.26e+01 1.09e+01
2048 1.90e+01 1.34e+01 1.09e+01 9.44e+00
8192 1.70e+01 1.20e+01 9.76e+00 8.43e+00
32768 1.55e+01 1.09e+01 8.90e+00 7.69e+00
131072 1.43e+01 1.01e+01 8.23e+00 7.11e+00
Table 4: maxK
(hK
hmin,K
)1/2for meshes of Example 5 with � = 10−3.
an a posteriori error estimator for anisotropic meshes. In this way, an adaptive
refinement could perform, with the aim of recognize the region of the domain
where the layers are, improving the quality of approximation in the process.
These are the subjects of ongoing work.
Acknowledgments
First author has been partially supported by CONICYT FONDECYT 1130158
and project AFB170001 of the PIA Program: Concurso Apoyo a Centros Cient́ı-
ficos y Tecnológicos de Excelencia con Financiamiento Basal, by project VRID-
Enlace No. 218.013.044-1.0, Universidad de Concepción, and by Centro de
Investigación en Ingenieŕıa Matemática (CI2MA), Universidad de Concepción.
Second author has been partially supported by CONICET-Argentina under
grant PID No. 14420140100027CO, by ANPCyT PICT No. 2014-1771, and
by Universidad de Buenos Aires under grant UBACyT No. 20020120100050.
Finally, third author has been partially supported by CONICYT FONDECYT
1160320 and project AFB170001 of the PIA Program: Concurso Apoyo a Cen-
tros Cient́ıficos y Tecnológicos de Excelencia con Financiamiento Basal, and
by Centro de Investigación en Ingenieŕıa Matemática (CI2MA), Universidad de
Concepción.
30
-
References
[1] B. Cockburn, Discontinuous Galerkin methods for convection-dominated
problems, in: High-order methods for computational physics, Vol. 9 of
Lect. Notes Comput. Sci. Eng., Springer, Berlin, 1999, pp. 69–224. doi:
10.1007/978-3-662-03882-6_2.
URL https://doi.org/10.1007/978-3-662-03882-6_2
[2] B. Cockburn, C.-W. Shu, Runge-Kutta discontinuous Galerkin methods
for convection-dominated problems, J. Sci. Comput. 16 (3) (2001) 173–
261. doi:10.1023/A:1012873910884.
URL https://doi.org/10.1023/A:1012873910884
[3] P. Castillo, B. Cockburn, D. Schötzau, C. Schwab, Optimal a priori error
estimates for the hp-version of the local discontinuous Galerkin method
for convection-diffusion problems, Math. Comp. 71 (238) (2002) 455–478.
doi:10.1090/S0025-5718-01-01317-5.
URL https://doi.org/10.1090/S0025-5718-01-01317-5
[4] B. Cockburn, C. Dawson, Some extensions of the local discontinuous
Galerkin method for convection-diffusion equations in multidimensions,
in: The mathematics of finite elements and applications, X, MAFELAP
1999 (Uxbridge), Elsevier, Oxford, 2000, pp. 225–238. doi:10.1016/
B978-008043568-8/50014-6.
URL https://doi.org/10.1016/B978-008043568-8/50014-6
[5] B. Cockburn, B. Dong, J. Guzmán, A superconvergent LDG-hybridizable
Galerkin method for second-order elliptic problems, Math. Comp. 77 (264)
(2008) 1887–1916. doi:10.1090/S0025-5718-08-02123-6.
URL https://doi.org/10.1090/S0025-5718-08-02123-6
[6] B. Cockburn, C.-W. Shu, The local discontinuous Galerkin method for
time-dependent convection-diffusion systems, SIAM J. Numer. Anal. 35 (6)
(1998) 2440–2463. doi:10.1137/S0036142997316712.
URL https://doi.org/10.1137/S0036142997316712
31
https://doi.org/10.1007/978-3-662-03882-6_2https://doi.org/10.1007/978-3-662-03882-6_2http://dx.doi.org/10.1007/978-3-662-03882-6_2http://dx.doi.org/10.1007/978-3-662-03882-6_2https://doi.org/10.1007/978-3-662-03882-6_2https://doi.org/10.1023/A:1012873910884https://doi.org/10.1023/A:1012873910884http://dx.doi.org/10.1023/A:1012873910884https://doi.org/10.1023/A:1012873910884https://doi.org/10.1090/S0025-5718-01-01317-5https://doi.org/10.1090/S0025-5718-01-01317-5https://doi.org/10.1090/S0025-5718-01-01317-5http://dx.doi.org/10.1090/S0025-5718-01-01317-5https://doi.org/10.1090/S0025-5718-01-01317-5https://doi.org/10.1016/B978-008043568-8/50014-6https://doi.org/10.1016/B978-008043568-8/50014-6http://dx.doi.org/10.1016/B978-008043568-8/50014-6http://dx.doi.org/10.1016/B978-008043568-8/50014-6https://doi.org/10.1016/B978-008043568-8/50014-6https://doi.org/10.1090/S0025-5718-08-02123-6https://doi.org/10.1090/S0025-5718-08-02123-6http://dx.doi.org/10.1090/S0025-5718-08-02123-6https://doi.org/10.1090/S0025-5718-08-02123-6https://doi.org/10.1137/S0036142997316712https://doi.org/10.1137/S0036142997316712http://dx.doi.org/10.1137/S0036142997316712https://doi.org/10.1137/S0036142997316712
-
[7] P. Houston, C. Schwab, E. Süli, Discontinuous hp-finite element methods
for advection-diffusion-reaction problems, SIAM J. Numer. Anal. 39 (6)
(2002) 2133–2163. doi:10.1137/S0036142900374111.
URL https://doi.org/10.1137/S0036142900374111
[8] B. Cockburn, B. Dong, An analysis of the minimal dissipation local discon-
tinuous Galerkin method for convection-diffusion problems, J. Sci. Comput.
32 (2) (2007) 233–262. doi:10.1007/s10915-007-9130-3.
URL https://doi.org/10.1007/s10915-007-9130-3
[9] P. Bochev, T. J. R. Hughes, G. Scovazzi, A multiscale discontinuous
Galerkin method, in: Large-scale scientific computing, Vol. 3743 of Lec-
ture Notes in Comput. Sci., Springer, Berlin, 2006, pp. 84–93. doi:
10.1007/11666806_8.
URL https://doi.org/10.1007/11666806_8
[10] A. Buffa, T. J. R. Hughes, G. Sangalli, Analysis of a multiscale discontin-
uous Galerkin method for convection-diffusion problems, SIAM J. Numer.
Anal. 44 (4) (2006) 1420–1440. doi:10.1137/050640382.
URL https://doi.org/10.1137/050640382
[11] T. J. R. Hughes, G. Scovazzi, P. B. Bochev, A. Buffa, A multiscale discon-
tinuous Galerkin method with the computational structure of a continuous
Galerkin method, Comput. Methods Appl. Mech. Engrg. 195 (19-22) (2006)
2761–2787. doi:10.1016/j.cma.2005.06.006.
URL https://doi.org/10.1016/j.cma.2005.06.006
[12] B. Ayuso, L. D. Marini, Discontinuous Galerkin methods for advection-
diffusion-reaction problems, SIAM J. Numer. Anal. 47 (2) (2009) 1391–
1420. doi:10.1137/080719583.
URL https://doi.org/10.1137/080719583
[13] H. Zarin, H.-G. Roos, Interior penalty discontinuous approximations of
convection-diffusion problems with parabolic layers, Numer. Math. 100 (4)
32
https://doi.org/10.1137/S0036142900374111https://doi.org/10.1137/S0036142900374111http://dx.doi.org/10.1137/S0036142900374111https://doi.org/10.1137/S0036142900374111https://doi.org/10.1007/s10915-007-9130-3https://doi.org/10.1007/s10915-007-9130-3http://dx.doi.org/10.1007/s10915-007-9130-3https://doi.org/10.1007/s10915-007-9130-3https://doi.org/10.1007/11666806_8https://doi.org/10.1007/11666806_8http://dx.doi.org/10.1007/11666806_8http://dx.doi.org/10.1007/11666806_8https://doi.org/10.1007/11666806_8https://doi.org/10.1137/050640382https://doi.org/10.1137/050640382http://dx.doi.org/10.1137/050640382https://doi.org/10.1137/050640382https://doi.org/10.1016/j.cma.2005.06.006https://doi.org/10.1016/j.cma.2005.06.006https://doi.org/10.1016/j.cma.2005.06.006http://dx.doi.org/10.1016/j.cma.2005.06.006https://doi.org/10.1016/j.cma.2005.06.006https://doi.org/10.1137/080719583https://doi.org/10.1137/080719583http://dx.doi.org/10.1137/080719583https://doi.org/10.1137/080719583https://doi.org/10.1007/s00211-005-0598-1https://doi.org/10.1007/s00211-005-0598-1
-
(2005) 735–759. doi:10.1007/s00211-005-0598-1.
URL https://doi.org/10.1007/s00211-005-0598-1
[14] C. E. Baumann, J. T. Oden, A discontinuous hp finite element method
for convection-diffusion problems, Comput. Methods Appl. Mech. Engrg.
175 (3-4) (1999) 311–341. doi:10.1016/S0045-7825(98)00359-4.
URL https://doi.org/10.1016/S0045-7825(98)00359-4
[15] H. Egger, J. Schöberl, A hybrid mixed discontinuous Galerkin finite-
element method for convection-diffusion problems, IMA J. Numer. Anal.
30 (4) (2010) 1206–1234. doi:10.1093/imanum/drn083.
URL https://doi.org/10.1093/imanum/drn083
[16] B. Cockburn, B. Dong, J. Guzmán, M. Restelli, R. Sacco, A hybridiz-
able discontinuous Galerkin method for steady-state convection-diffusion-
reaction problems, SIAM J. Sci. Comput. 31 (5) (2009) 3827–3846. doi:
10.1137/080728810.
URL https://doi.org/10.1137/080728810
[17] N. C. Nguyen, J. Peraire, B. Cockburn, An implicit high-order hybridizable
discontinuous Galerkin method for linear convection-diffusion equations, J.
Comput. Phys. 228 (9) (2009) 3232–3254. doi:10.1016/j.jcp.2009.01.
030.
URL https://doi.org/10.1016/j.jcp.2009.01.030
[18] N. C. Nguyen, J. Peraire, B. Cockburn, An implicit high-order hybridizable
discontinuous Galerkin method for nonlinear convection-diffusion equa-
tions, J. Comput. Phys. 228 (23) (2009) 8841–8855. doi:10.1016/j.jcp.
2009.08.030.
URL https://doi.org/10.1016/j.jcp.2009.08.030
[19] G. Fu, W. Qiu, W. Zhang, An analysis of HDG methods for convection-
dominated diffusion problems, ESAIM Math. Model. Numer. Anal. 49 (1)
(2015) 225–256. doi:10.1051/m2an/2014032.
URL https://doi.org/10.1051/m2an/2014032
33
http://dx.doi.org/10.1007/s00211-005-0598-1https://doi.org/10.1007/s00211-005-0598-1https://doi.org/10.1016/S0045-7825(98)00359-4https://doi.org/10.1016/S0045-7825(98)00359-4http://dx.doi.org/10.1016/S0045-7825(98)00359-4https://doi.org/10.1016/S0045-7825(98)00359-4https://doi.org/10.1093/imanum/drn083https://doi.org/10.1093/imanum/drn083http://dx.doi.org/10.1093/imanum/drn083https://doi.org/10.1093/imanum/drn083https://doi.org/10.1137/080728810https://doi.org/10.1137/080728810https://doi.org/10.1137/080728810http://dx.doi.org/10.1137/080728810http://dx.doi.org/10.1137/080728810https://doi.org/10.1137/080728810https://doi.org/10.1016/j.jcp.2009.01.030https://doi.org/10.1016/j.jcp.2009.01.030http://dx.doi.org/10.1016/j.jcp.2009.01.030http://dx.doi.org/10.1016/j.jcp.2009.01.030https://doi.org/10.1016/j.jcp.2009.01.030https://doi.org/10.1016/j.jcp.2009.08.030https://doi.org/10.1016/j.jcp.2009.08.030https://doi.org/10.1016/j.jcp.2009.08.030http://dx.doi.org/10.1016/j.jcp.2009.08.030http://dx.doi.org/10.1016/j.jcp.2009.08.030https://doi.org/10.1016/j.jcp.2009.08.030https://doi.org/10.1051/m2an/2014032https://doi.org/10.1051/m2an/2014032http://dx.doi.org/10.1051/m2an/2014032https://doi.org/10.1051/m2an/2014032
-
[20] Y. Chen, B. Cockburn, Analysis of variable-degree HDG methods for
convection-diffusion equations. Part I: general nonconforming meshes, IMA
J. Numer. Anal. 32 (4) (2012) 1267–1293. doi:10.1093/imanum/drr058.
URL https://doi.org/10.1093/imanum/drr058
[21] Y. Chen, B. Cockburn, Analysis of variable-degree HDG methods
for convection-diffusion equations. Part II: Semimatching nonconform-
ing meshes, Math. Comp. 83 (285) (2014) 87–111. doi:10.1090/
S0025-5718-2013-02711-1.
URL https://doi.org/10.1090/S0025-5718-2013-02711-1
[22] B. Cockburn, J. Gopalakrishnan, F.-J. Sayas, A projection-based er-
ror analysis of HDG methods, Math. Comp. 79 (271) (2010) 1351–1367.
doi:10.1090/S0025-5718-10-02334-3.
URL https://doi.org/10.1090/S0025-5718-10-02334-3
[23] W. Eckhaus, Boundary layers in linear elliptic singular perturbation prob-
lems, SIAM Rev. 14 (1972) 225–270. doi:10.1137/1014030.
URL https://doi.org/10.1137/1014030
[24] H. Goering, A. Felgenhauer, G. Lube, H.-G. Roos, L. Tobiska, Singu-
larly perturbed differential equations, Vol. 13 of Mathematical Research,
Akademie-Verlag, Berlin, 1983.
[25] H. Chen, J. Li, W. Qiu, Robust a posteriori error estimates for HDG
method for convection-diffusion equations, IMA J. Numer. Anal. 36 (1)
(2016) 437–462. doi:10.1093/imanum/drv009.
URL https://doi.org/10.1093/imanum/drv009
[26] H. Zhu, Z. Zhang, Uniform convergence of the LDG method for a singu-
larly perturbed problem with the exponential boundary layer, Math. Comp.
83 (286) (2014) 635–663. doi:10.1090/S0025-5718-2013-02736-6.
URL https://doi.org/10.1090/S0025-5718-2013-02736-6
34
https://doi.org/10.1093/imanum/drr058https://doi.org/10.1093/imanum/drr058http://dx.doi.org/10.1093/imanum/drr058https://doi.org/10.1093/imanum/drr058https://doi.org/10.1090/S0025-5718-2013-02711-1https://doi.org/10.1090/S0025-5718-2013-02711-1https://doi.org/10.1090/S0025-5718-2013-02711-1http://dx.doi.org/10.1090/S0025-5718-2013-02711-1http://dx.doi.org/10.1090/S0025-5718-2013-02711-1https://doi.org/10.1090/S0025-5718-2013-02711-1https://doi.org/10.1090/S0025-5718-10-02334-3https://doi.org/10.1090/S0025-5718-10-02334-3http://dx.doi.org/10.1090/S0025-5718-10-02334-3https://doi.org/10.1090/S0025-5718-10-02334-3https://doi.org/10.1137/1014030https://doi.org/10.1137/1014030http://dx.doi.org/10.1137/1014030https://doi.org/10.1137/1014030https://doi.org/10.1093/imanum/drv009https://doi.org/10.1093/imanum/drv009http://dx.doi.org/10.1093/imanum/drv009https://doi.org/10.1093/imanum/drv009https://doi.org/10.1090/S0025-5718-2013-02736-6https://doi.org/10.1090/S0025-5718-2013-02736-6http://dx.doi.org/10.1090/S0025-5718-2013-02736-6https://doi.org/10.1090/S0025-5718-2013-02736-6
-
[27] T. Apel, Anisotropic finite elements: local estimates and applications, Ad-
vances in Numerical Mathematics, B. G. Teubner, Stuttgart, 1999.
[28] T. Dupont, R. Scott, Polynomial approximation of functions in Sobolev
spaces, Math. Comp. 34 (150) (1980) 441–463. doi:10.2307/2006095.
URL https://doi.org/10.2307/2006095
[29] G. Kunert, Error estimation for anisotropic tetrahedral and triangular finite
element meshes, Tech. Rep. SFB393/97-16, Fakultät Mathematik, Technis-
che Universität Chemnitz, Chemnitz, Germany (1997).
[30] F. Hecht, BAMG: bidimensional anisotropic mesh generator. User guide
(1998).
[31] B. Cockburn, W. Qiu, M. Solano, A priori error analysis for HDG
methods using extensions from subdomains to achieve boundary con-
formity, Math. Comp. 83 (286) (2014) 665–699. doi:10.1090/
S0025-5718-2013-02747-0.
URL https://doi.org/10.1090/S0025-5718-2013-02747-0
[32] H.-G. Roos, M. Stynes, L. Tobiska, Robust numerical methods for sin-
gularly perturbed differential equations, 2nd Edition, Vol. 24 of Springer
Series in Computational Mathematics, Springer-Verlag, Berlin, 2008,
convection-diffusion-reaction and flow problems.
35
https://doi.org/10.2307/2006095https://doi.org/10.2307/2006095http://dx.doi.org/10.2307/2006095https://doi.org/10.2307/2006095https://doi.org/10.1090/S0025-5718-2013-02747-0https://doi.org/10.1090/S0025-5718-2013-02747-0https://doi.org/10.1090/S0025-5718-2013-02747-0http://dx.doi.org/10.1090/S0025-5718-2013-02747-0http://dx.doi.org/10.1090/S0025-5718-2013-02747-0https://doi.org/10.1090/S0025-5718-2013-02747-0
-
k N eq e.o.c eu e.o.c eσ e.o.c eû e.o.c
0 32 1.98e-01 −− 4.88e-01 −− 6.89e-01 −− 7.07e+00 −−
128 7.44e-02 0.41 3.40e-01 −0.48 4.81e-01 −0.48 5.13e+00 −0.54
512 4.60e-02 0.11 1.79e-01 0.34 2.54e-01 0.34 3.73e+00 −0.12
2048 3.46e-02 −0.01 9.31e-02 0.53 1.32e-01 0.53 2.68e+00 0.06
8192 2.87e-02 −0.05 4.88e-02 0.61 6.90e-02 0.61 1.92e+00 0.16
32768 2.54e-02 −0.09 2.65e-02 0.62 3.74e-02 0.62 1.38e+00 0.21
131072 2.28e-02 −0.07 1.53e-02 0.57 2.16e-02 0.57 1.00e+00 0.24
1 32 3.25e-02 −− 2.24e-01 −− 3.17e-01 −− 1.74e+00 −−
128 1.37e-02 0.24 4.26e-02 1.39 6.03e-02 1.39 7.60e-01 0.19
512 5.66e-03 0.69 9.57e-03 1.57 1.35e-02 1.57 3.44e-01 0.56
2048 2.43e-03 0.80 2.24e-03 1.68 3.17e-03 1.68 1.53e-01 0.75
8192 1.05e-03 0.88 5.33e-04 1.75 7.49e-04 1.76 6.63e-02 0.88
32768 4.80e-04 0.87 1.25e-04 1.83 1.71e-04 1.86 2.80e-02 0.98
131072 2.48e-04 0.72 2.76e-05 1.95 3.49e-05 2.07 1.17e-02 1.03
2 32 1.65e-02 −− 2.67e-02 −− 3.81e-02 −− 3.42e-01 −−
128 3.00e-03 1.46 2.54e-03 2.40 3.63e-03 2.39 1.18e-01 0.54
512 6.95e-04 1.52 2.57e-04 2.71 3.71e-04 2.70 3.94e-02 0.99
2048 1.74e-04 1.58 2.92e-05 2.72 4.34e-05 2.68 1.18e-02 1.32
8192 4.03e-05 1.78 3.77e-06 2.63 5.91e-06 2.55 3.21e-03 1.55
32768 8.72e-06 1.94 5.25e-07 2.58 8.73e-07 2.49 8.16e-04 1.71
131072 2.04e-06 1.87 7.99e-08 2.49 1.41e-07 2.41 1.99e-04 1.81
3 32 9.26e-03 −− 3.91e-03 −− 5.68e-03 −− 7.45e-02 −−
128 9.70e-04 2.25 2.18e-04 3.16 3.18e-04 3.15 1.82e-02 1.03
512 1.62e-04 1.99 1.54e-05 3.23 2.38e-05 3.15 4.98e-03 1.28
2048 2.68e-05 2.18 1.37e-06 3.07 2.61e-06 2.77 1.03e-03 1.86
8192 3.66e-06 2.55 1.63e-07 2.75 3.38e-07 2.62 1.77e-04 2.22
32768 4.16e-07 2.87 1.96e-08 2.78 4.09e-08 2.78 2.70e-05 2.44
131072 3.75e-08 3.24 2.20e-09 2.93 4.48e-09 2.97 3.85e-06 2.59
Table 5: History of convergence of Example 5. � = 10−3.
36
-
N k = 0 k = 1 k = 2 k = 3
32 3.80e+04 2.69e+04 2.19e+04 1.90e+04
128 2.69e+04 1.90e+04 1.55e+04 1.34e+04
512 2.19e+04 1.55e+04 1.27e+04 1.10e+04
2048 1.90e+04 1.34e+04 1.10e+04 9.50e+03
8192 1.70e+04 1.20e+04 9.81e+03 8.49e+03
32768 1.55e+04 1.10e+04 8.95e+03 7.75e+03
131072 1.44e+04 1.02e+04 8.29e+03 7.18e+03
Table 6: maxK∈T
(hK
hmin,K
)1/2for meshes of Example 5 with � = 10−9.
37
-
k N eq e.o.c eu e.o.c eσ e.o.c eû e.o.c
0 32 1.94e-01 −− 4.88e-01 −− 6.90e-01 −− 6.28e+03 −−
128 6.55e-02 0.56 3.39e-01 −0.47 4.80e-01 −0.47 4.36e+03 −0.47
512 3.57e-02 0.29 1.77e-01 0.35 2.50e-01 0.35 3.12e+03 −0.11
2048 2.25e-02 0.25 9.00e-02 0.56 1.27e-01 0.56 2.23e+03 0.07
8192 1.39e-02 0.37 4.53e-02 0.67 6.41e-02 0.67 1.59e+03 0.17
32768 8.29e-03 0.48 2.28e-02 0.73 3.22e-02 0.73 1.13e+03 0.23
131072 4.80e-03 0.57 1.14e-02 0.77 1.61e-02 0.77 8.03e+02 0.27
1 32 1.15e-02 −− 2.25e-01 −− 3.19e-01 −− 1.29e+03 −−
128 7.59e-03 −0.39 4.32e-02 1.38 6.11e-02 1.38 6.75e+02 −0.07
512 3.88e-03 0.38 9.85e-03 1.55 1.39e-02 1.55 3.26e+02 0.47
2048 1.66e-03 0.81 2.36e-03 1.65 3.34e-03 1.65 1.48e+02 0.72
8192 6.42e-04 1.05 5.77e-04 1.71 8.17e-04 1.71 6.43e+01 0.88
32768 2.33e-04 1.20 1.43e-04 1.75 2.02e-04 1.75 2.69e+01 0.99
131072 8.01e-05 1.32 3.55e-05 1.78 5.02e-05 1.78 1.10e+01 1.07
2 32 6.41e-03 −− 2.64e-02 −− 3.74e-02 −− 2.00e+02 −−
128 1.56e-03 1.04 2.51e-03 2.40 3.54e-03 2.40 1.05e+02 −0.07
512 5.27e-04 0.98 2.55e-04 2.71 3.61e-04 2.71 3.84e+01 0.86
2048 1.42e-04 1.48 2.85e-05 2.75 4.03e-05 2.75 1.18e+01 1.29
8192 3.36e-05 1.76 3.39e-06 2.75 4.80e-06 2.75 3.20e+00 1.56
32768 7.06e-06 1.99 4.16e-07 2.77 5.88e-07 2.77 8.05e-01 1.73
131072 1.40e-06 2.12 5.15e-08 2.79 7.28e-08 2.79 1.92e-01 1.85
3 32 7.35e-04 −− 3.92e-03 −− 5.54e-03 −− 3.49e+01 −−
128 5.30e-04 −0.53 2.20e-04 3.16 3.11e-04 3.16 1.78e+01 −0.03
512 1.47e-04 1.27 1.50e-05 3.29 2.12e-05 3.29 5.00e+00 1.25
2048 2.64e-05 2.06 9.57e-07 3.55 1.35e-06 3.55 1.04e+00 1.86
8192 3.73e-06 2.50 6.04e-08 3.66 8.54e-08 3.66 1.77e-01 2.22
32768 4.73e-07 2.71 3.79e-09 3.73 5.36e-09 3.73 2.70e-02 2.46
131072 8.40e-08 2.27 2.37e-10 3.77 3.36e-10 3.77 3.82e-03 2.59
Table 7: History of convergence of Example 5. � = 10−9.
38
IntroductionConvection–diffusion problemAn anisotropic a priori error analysisNumerical resultsUnstructured meshesBoundary layersInterior layerNon constant convection
Shishkin meshes
Conclusions and final comments